Physics-aware graph neural ODE for sensor data imputation in converter transformer monitoring systems

Online monitoring systems for converter transformers, comprising IoT sensor networks with 21 multi-point deployed sensors, frequently suffer from missing values due to sensor failures and communication malfunctions, affecting grid reliability. Traditional methods struggle to simultaneously model multi-timescale characteristics and cross-physical-domain coupling relationships of different parameter types. This paper proposes Physics-aware Type-differentiated Graph Neural Ordinary Differential Equation (PT-GNODE) for imputing missing values in converter transformer monitoring data. PT-GNODE combines the spatial modeling capabilities of graph neural networks with continuous temporal evolution characteristics of neural ordinary differential equations. The model precisely characterizes electrical-temperature-gas cascading physical processes through parameter type-aware ODE function systems, employs dynamic graph learning mechanisms to adapt to operating condition variations, and embeds physical constraints including temperature gradients, gas balance, and current ranges. Additionally, multi-scale feature enhancement mechanisms effectively utilize spatiotemporal contextual information to improve imputation accuracy. Experiments on 36 months of actual operational data from an 800 kV converter station demonstrate that PT-GNODE outperforms baseline methods across different missing rates, achieving coefficients of determination of 0.956, 0.957, and 0.937 for electrical, temperature, and gas parameters respectively even under severe missing conditions, exhibiting exceptional robustness and physical consistency.The proposed method is suitable for deployment in industrial IoT monitoring systems with real-time processing requirements.